A combo of CI unblocks via PRs #2106 #2134 #2141 #2142 #2146 #2145
Conversation
This is necessary because: * `pypy-3.8` is EOL * `pypy-3.8` has flaky SEGFAULTs on import [[1]] [[2]] [[3]] due to a bug in their GC that is fixed in `pypy-3.9` [1]: https://github.com/jazzband/pip-tools/actions/runs/12162197242/job/33918558133?pr=2106#step:8:59 [2]: pytest-dev/pytest#11771 (comment) [3]: https://pypy.org/posts/2024/03/fixing-bug-incremental-gc.html
This better reflects that the pip version being pulled in is the lowest tested/supported.
This is meant to be separate from "latest" and is pinned to the known working released version of pip. It will run on merges to `main` and in pull requests. And the cron runs will additionally test the latest version on PyPI and the `main` branch of the pip repo.
This fixes a bug that build deps compilation would get the latest version of an unconstrained build requirements list, not taking into account the restricted/requested one. The regression test is implemented against `setuptools < 70.1.0` which is known to inject a dependency on `wheel` (newer `setuptools` vendor it). The idea is that `pyproject.toml` does not have an upper bound for `setuptools` but the CLI arg does. And when this works correctly, the `wheel` entry will be included into the resolved output. Cleaning up `PIP_CONSTRAINT` is implemented manually due to the corner case of a permission error on Windows when accessing a file that we hold a file descriptor to from another subprocess[[1]]. It can be further simplified once the lowest Python version `pip-tools` supports is 3.12 by replacing `delete=False` with `delete_on_close=False` in the `tempfile.NamedTemporaryFile()` context manager initializer. [1]: https://docs.python.org/3/library/tempfile.html#tempfile.NamedTemporaryFile Co-authored-by: Sviatoslav Sydorenko <[email protected]>
The `pre-commit.ci` service dropped support for Python 3.8 as it's gone EOL. This change makes the check runnable again. It is suboptimal as we haven't yet dropped support for Python 3.8. jazzband#2144 will address the underlying issue more thoroughly. Resolves jazzband#2133. Co-Authored-By: Sviatoslav Sydorenko <[email protected]>
webknjaz
added
bug
webknjaz
force-pushed
the
maintenance/reproducible-ci-fixes-pr2106-pr2141-pr2142
branch
from
2741e7c to
51ec5ed
Compare
webknjaz
changed the title
|
@hugovk @chrysle @Mr-Sunglasses @macro1 as soon as somebody approves this, everybody should get unblocked. This contains a true/natural merge commit of individual branches in 4 PRs. They are the same commits (hashes and everything) so when merging this, GitHub is supposed to auto-detect that it can mark those other PRs as merged automatically. |
Basically, it turned out that the variable we relied on was always unavailable and so CI was always evaluating it the same way.
No, merge commits don't contain duplicates by their nature. We mustn't rebase, this is done on purpose so that the actual flow of history is preserved. If you rebase, you'll basically disregard the original commits, replacing them with copies in a weird order and no link to reality. This is an intentional “octopus” merge in hopes that the merge queues mechanism would just fast-forward |
|
The PyPy failure is rather weird. I saw similar problems on Windows but CPython. And I thought that was gone. I'll have to debug this last thing. |
So this is actually a bug in PyPy. Also seen in tox: tox-dev/tox#3284. PyPy maintainers recommend testing under PyPy 3.10 instead: pypy/pypy#4958 (comment). So I'm going to add another PR with that bump. |
This is necessary because: * `pypy-3.8` is EOL * `pypy-3.8` has flaky SEGFAULTs on import [[1]] [[2]] [[3]] due to a bug in their GC that is fixed in `pypy-3.9` * `pypy-3.9` has a bug that does not have a backport with the fix [[4]]. Furthermore, PyPy maintainers recommend using 3.10 [[5]]. [1]: https://github.com/jazzband/pip-tools/actions/runs/12162197242/job/33918558133?pr=2106#step:8:59 [2]: pytest-dev/pytest#11771 (comment) [3]: https://pypy.org/posts/2024/03/fixing-bug-incremental-gc.html [4]: tox-dev/tox#3284 [5]: pypy/pypy#4958 (comment)
webknjaz
changed the title
webknjaz
force-pushed
the
maintenance/reproducible-ci-fixes-pr2106-pr2141-pr2142
branch
from
51ec5ed to
fb0f69b
Compare
|
I was considering pinning to But this implies that pypy-3.9 is EOL so I won't bother: pypy/pypy#4958 (comment). |
Nope, it still created another merge commit. Though, it's fine with me since it still is supposed to work, backtracking the PRs to mark as merged. |
|
Thanks for picking up this one and stabilising CI! 🙌 |
This is necessary to merge as a combo because of how the branch protection is configured.
Contributor checklist
Maintainer checklist
backwards incompatible,feature,enhancement,deprecation,bug,dependency,docsorskip-changelogas they determine changelog listing.